Agentic Context Engineering (ACE): Evolving Contexts for Self-Improving Language Models

> arXiv:2510.04618 | Code: https://github.com/ace-agent/ace

Introduction

Large Language Model applications increasingly rely on context adaptation rather than weight updates. Current approaches suffer from two critical limitations:

• Brevity bias: over-prioritizing concise summaries at the expense of detailed domain insights
• Context collapse: iterative rewriting erodes details over time, leading to performance drops

ACE treats contexts as evolving playbooks that accumulate, refine, and organize strategies through a modular process.

Three-Role Architecture

1. Generator — produces reasoning trajectories for new queries, surfacing effective strategies and pitfalls 2. Reflector — critiques generated traces, distilling insights from successes and errors 3. Curator — synthesizes insights into structured "delta entries" and integrates them into the existing context

Key Innovations

Incremental Delta Updates

• Contexts are represented as structured, itemized "bullets" with metadata and content
• Small, localized edits preserve prior knowledge while accumulating new insights
• Non-LLM logic performs deterministic merging, de-duplication, and pruning

Grow-and-Refine Mechanism

• Balances context expansion with periodic refinement
• Maintains relevance and prevents unbounded growth
• Enables efficient, parallel merging crucial for scalability

Performance Results

ACE consistently outperforms strong baselines across agent and domain-specific benchmarks:

• +10.6% on agent tasks (AppWorld)
• +8.6% on financial analysis (FiNER + XBRL)
• Matches a top-ranked production-level agent on the AppWorld leaderboard using a smaller open-source model

Efficiency Gains

| Metric | Offline vs GEPA | Online vs Dynamic Cheatsheet | | --- | --- | --- | | Latency reduction | 82.3% | 91.5% | | Rollout/token cost reduction | 75.1% | 83.6% |

ACE also adapts effectively without labeled supervision by leveraging natural execution feedback.

Implications

• Enables scalable, efficient, and self-improving LLM systems with low overhead
• Provides interpretable contexts and lower overhead compared to fine-tuning
• Offers a flexible approach for online and continuous learning
• Particularly valuable for specialized domains and long-context applications